Everything indie hackers need to know about OpenAI's new model

Chances are you've seen that OpenAI released their new model "o1" this week.

Now, it's time to learn everything you need to know about what OpenAI founder Sam Altman has called the beginning of a new paradigm: AI that can do complex reasoning.

What is o1?

o1 — previously codenamed "Strawberry" — is a new series of AI models from OpenAI. It's the latest release in a long line of models from the company.

The o1 models are designed to spend extra time thinking before responding to your requests. But in return, they can reason through complex, multi-step tasks, one step at a time. This allows them to solve trickier problems than existing models.

There are two o1 models, both of which are accessible through ChatGPT:

• o1-preview, the most powerful new o1 model.

• o1-mini, a faster, 80% cheaper model that's especially good at coding and other applications that need reasoning, but not broad world knowledge. 

Overall, the o1 models excel in fields like science, math, and coding.

There are impressive examples showing that o1 is much better at things like doing your physics homework, translating a corrupted sentence, coding a video game from a prompt, and perhaps most impressively, answering how many r's are in "strawberry."

How does it work behind the hood?

o1 uses a mix of chain-of-thought and reinforcement learning to reason through problems. At a high level, this just means that it breaks down problems into simple steps, and then goes through those steps until it reaches the right answer. If for any reason it hits a snag, it simply tries a different approach.

When you think about this, it's really hard to wrap your head around it. How in the world can an AI know if what it is doing is correct?

Unfortunately, OpenAI hasn't published their research, but those in the know are hypothesizing that there's a separate critic model scoring outputs with a heuristic. Hopefully we'll learn the full truth soon.

How good is it?

Really, really good on the benchmarks. Check out how it compares to OpenAI's previous model, GPT-4o:

There are even more detailed benchmarks on OpenAI's page about learning to reason with LLMs.

However, we shouldn't get too excited just yet.

This is the beginning of a new frontier, which means there's still a lot of room for improvement. On X, Altman described o1 as "[seeming] more impressive on first use than it does after you spend more time with it."

Why does o1 matter?

Up to now, reasoning was impossible for AI models. They were mere "stochastic parrots" trained on immense amounts of data that enabled them to predict the correct answer to simple questions. Now, they're capable of thinking through questions in much the same way that a human does.

That's a whole new level unlocked, which OpenAI itself acknowledges. Hence the name o1, instead of GPT-5.

Noam Brown, who researches reasoning at OpenAI, had this to say:

“o1 thinks for seconds, but we aim for future versions to think for hours, days, even weeks. Inference costs will be higher, but what cost would you pay for a new cancer drug?”

What should you do with o1?

First, play around with it.

You'll need a ChatGPT Plus subscription, which costs $20/month, or a subscription for your team, which costs $25/user/month. Then all you need to do is log into ChatGPT and select the o1 model of your choice from the menu at the top of the screen:

Odds are that it can help you with something. So get in there and start using it.

By doing so, you'll learn it's strengths and weaknesses, and thus, what you can build with it. Countless startups were born from GPT-4, and this is the biggest AI advance since then.

Now is not the time to stay on the sidelines.